Construction of an ear disease classification system based on lesion characteristics in high-resolution CT of the temporal bone

Abstract

The present invention discloses an ear disease classification system based on lesion features in high-resolution CT of the temporal bone. The system is characterized by comprising the following steps: annotating temporal bone lesions, constructing a disease classification system based on a lesion feature library and notes, converting the data set format and dividing it into a training set and a test set, recording model pre-training weights and setting hyperparameters, loading classification data and performing data enhancement, inferring the model to obtain classification results and calculating training loss, backpropagating the training loss and iterating, determining whether convergence has occurred, and repeating the calculation if not; if convergence has occurred, inferring the model on the test set and verifying the lesion classification results, and finally having a professional doctor evaluate the classification results and apply them to clinical diagnosis. The present invention is based on identifying multiple ear disease features in high-resolution CT of the temporal bone and then further constructing an ear disease classification system. It is a classification system with strong clinical interpretability based on strongly supervised learning, effectively improving clinical diagnostic efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of medical imaging technology, and in particular relates to constructing an ear disease classification system based on lesion characteristics in temporal bone high-resolution CT. Background Art

[0002] Existing deep learning studies all use weakly supervised learning to construct models [1]. That is, only the disease labels of high-resolution CT of the temporal bone are input, but the fine lesion features of the high-resolution CT of the temporal bone are not annotated and input into the deep learning model. As a result, the interpretability of the constructed deep learning model is relatively limited, and there is room for improvement in the diagnostic accuracy. Therefore, it is difficult to convince clinical physicians to trust such models, which in turn limits their clinical use.

[0003] Many existing models use some key layers in temporal bone high-resolution CT[2] rather than all layer files of temporal bone high-resolution CT as input data to build deep learning models. However, many diseases have multiple characteristics that exist in multiple anatomical parts (multiple layers). Therefore, using only some layers of temporal bone high-resolution CT to diagnose diseases has data input limitations, and it is difficult for clinicians to believe that a robust diagnostic model can be built based on only some layers.

[0004] In addition, more studies have focused on the segmentation of multiple anatomical structures in high-resolution CT of the temporal bone [3]. The purpose is to build a model with accurate anatomical structure segmentation, but the segmentation has not been used for disease classification and diagnosis. Summary of the Invention

[0005] The purpose of the present invention is to provide an ear disease classification system based on lesion features in high-resolution temporal bone CT scans. By extracting lesion features from high-resolution temporal bone CT scans in ear cases, a library of lesion features for automatic identification of lesions on high-resolution temporal bone CT scans is constructed, and an interpretable ear disease diagnosis model is trained based on the lesion features. Based on the identification of lesion features on high-resolution temporal bone CT scans, ear diseases are further classified. To achieve this purpose, the technical solution adopted by the present invention is as follows:

[0006] The construction of an ear disease classification system based on the lesion characteristics of temporal bone high-resolution CT includes the following steps:

[0007] Step S201: The clinical physician team is trained on the use of annotation software, the imaging lesion features of temporal bone high-resolution CT, and the feature annotation standards. The annotation content of temporal bone high-resolution CT includes the disease name and the lesion features of each disease in the temporal bone high-resolution CT, thereby constructing a temporal bone high-resolution CT imaging lesion feature library;

[0008] Step S202: constructing a classification dataset of ear diseases based on the imaging lesion feature library and disease name labels of the temporal bone high-resolution CT imaging lesions;

[0009] Step S203: Convert the classification dataset format of the constructed temporal bone high-resolution CT-based imaging lesion classification dataset according to the lesion category, and divide the images and label files in the training set and test set into a ratio of 7:3, that is, 70% of the images are used for neural network model training, and 30% of the images are used for neural network model effect verification and evaluation;

[0010] Step S204: Load the pre-trained weights of the neural network-based classification model and set hyperparameters, including learning rate, batch data size, input image size, weight decay, and random seed parameters;

[0011] Step S205: Load the divided temporal bone high-resolution CT-based imaging lesion classification dataset and perform a certain degree of data augmentation on the dataset, including random cropping, upside-down flipping, and left-right flipping strategies, to accelerate model convergence and improve the robustness and generalization ability of the model;

[0012] Step S206: Input the loaded data into the loaded neural network-based temporal bone high-resolution CT imaging lesion classification model; temporal bone high-resolution CT imaging lesion classification is a binary classification problem, and the model inference obtains a classification result x=[x1,x2], where x1 is the predicted score of a normal CT image and x2 is the predicted score of an abnormal CT image; the label of the CT image is y=Y, Y∈(0,1), and the label is converted into a hot encoding form, i.e., y=[y1,y2]; the classification results x and y obtained by model inference are input into the cross entropy loss function to calculate the training loss, and the calculation formula is loss=-(y1logx1+y2logx2);

[0013] Step S207: The training loss calculated in step S206 is back-propagated to obtain the optimized gradient of each weight of the model, thereby optimizing the model parameters.

[0014] Step S208: Determine whether the model has converged. If it has converged to a near global optimum, save the model and proceed to step S209. If it has not converged, return to step S206 and iterate and optimize the model.

[0015] Step S209: performing inference on the trained and saved neural network classification model on the divided temporal bone high-resolution CT-based imaging lesion classification test set, outputting the lesion classification results, and performing visual verification;

[0016] Step S2010: A team of professional clinical physicians evaluates the output lesion classification results to verify the accuracy and generalization ability of the model, and further applies the neural network-based temporal bone high-resolution CT imaging lesion classification model to the clinical diagnosis of lesions.

[0017] Furthermore, the specific steps of the random cropping are as follows: first, the input image is scaled to 512x512 size, and then a 384x384 image block is cropped at an arbitrary position as the input of the model; in the data enhancement of upside-down and left-right flipping, the flipping probability is set to 0.5.

[0018] Furthermore, the pathological features include air-fluid level in secretory otitis media, enlargement of the tympanic sinus entrance in cholesteatoma, enlargement of the Prussak space, displacement and erosion of the ossicular chain, enlargement of the mastoid sinus filled with soft tissue and bone destruction, destruction of the scutum, destruction of the facial nerve canal and destruction of the lateral semicircular canal.

[0019] Furthermore, the annotated disease names include external auditory canal cholesteatoma, middle ear cholesteatoma, suppurative otitis media, cholesterol granuloma, secretory otitis media, temporal bone cancer, cochlear malformation, glomus jugularis tumor and acoustic neuroma.

[0020] Beneficial effects of the present invention:

[0021] 1. The disease classification logic of the present invention is different from that of a classification system based on weakly supervised learning. The present invention is based on identifying the characteristics of multiple ear diseases in high-resolution CT of the temporal bone and then further constructing a classification system for ear diseases. This is a classification system with strong clinical interpretability based on strongly supervised learning, which effectively improves clinical diagnostic efficiency.

[0022] 2. The present invention identifies multiple features of multiple ear diseases in multiple temporal bone high-resolution CT scans and constructs a temporal bone high-resolution CT lesion feature library for common ear diseases.

[0023] 3. The present invention organically combines various features of various ear diseases in various temporal bone high-resolution CT images to construct a classification system for ear diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flowchart provided for an embodiment of the present invention. Implementation Method

[0025] To facilitate understanding of the present invention, the present invention will be described more fully below with reference to the accompanying drawings. The accompanying drawings illustrate preferred embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present disclosure.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0027] The technical solution of this patent is further described in detail below in conjunction with specific implementation methods.

[0028] like Figure 1 As shown, an embodiment of the present invention provides an ear disease classification system based on the lesion characteristics in the temporal bone high-resolution CT, including the following steps: screening the temporal bone high-resolution CT to exclude other low-quality temporal bone high-resolution CT with artifacts.

[0029] The data was divided into training and test sets, and a team of clinical physicians was trained in the annotation software, the imaging features of temporal bone high-resolution CT scans, and the feature annotation standards. The annotations included the disease name and the temporal bone features contained in the high-resolution CT scans. A team of 10 clinical physicians performed the first round of annotation, followed by the first round of review and revision by three attending physicians or above, and the second round of review by three associate chief physicians. Any discrepancies between the first and second rounds of review and revision were discussed until a consensus was reached.

[0030] The dataset was cleaned to remove incorrect and missed annotations. Each lesion feature included at least 20 high-resolution CT images of the temporal bone.

[0031] Step S201: The clinical team is trained in the use of annotation software, the features of temporal bone HRCT imaging lesions, and the annotation standards. Temporal bone HRCT annotations include disease names, including external auditory canal cholesteatoma, middle ear cholesteatoma, suppurative otitis media, cholesterol granuloma, secretory otitis media, temporal bone cancer, cochlear malformation, glomus jugularis tumor, and acoustic neuroma. The lesion features of each disease on temporal bone HRCT are also included, such as the air-fluid level in secretory otitis media, enlarged tympanic sinus entrance, enlarged Prussak space, displacement and erosion of the ossicular chain, enlarged mastoid sinus with soft tissue filling and bone destruction, scutellum destruction, facial nerve canal destruction, and lateral semicircular canal destruction. This constructs a temporal bone HRCT imaging lesion feature library.

[0032] Step S202: constructing a classification dataset of ear diseases based on the imaging lesion features of the temporal bone high-resolution CT according to the imaging lesion feature library and disease name labels of the temporal bone high-resolution CT.

[0033] Step S203: For the constructed imaging lesion classification dataset based on high-resolution CT of the temporal bone, the classification dataset format is converted according to the lesion category, and the images and label files in the training set and the test set are divided in a ratio of 7:3, that is, 70% of the images are used for neural network model training, and 30% of the images are used for neural network model effect verification and evaluation.

[0034] Step S204: Load the pre-trained weights of the neural network-based temporal bone high-resolution CT imaging lesion classification model and set the hyperparameters: the learning rate is set to 0.000045, the batch size is set to 8, the input image is uniformly scaled to 384x384, the weight decay is set to 0.01, and the random seed is set to 42.

[0035] Step S205: Load the partitioned temporal bone high-resolution CT-based imaging lesion classification dataset and perform a certain degree of data augmentation on the dataset, including random cropping, upside-down flipping, and left-right flipping strategies, to accelerate model convergence and improve model robustness and generalization. The specific steps for random cropping are as follows: first scale the input image to 512x512, then crop a 384x384 image block at an arbitrary location as the model input. For the upside-down and left-right flipping data augmentation, the flip probability is set to 0.5.

[0036] Step S206: Input the loaded data into the loaded neural network-based temporal bone high-resolution CT imaging lesion classification model; temporal bone high-resolution CT imaging lesion classification is a binary classification problem, and the model inference obtains the classification result x=[x1,x2], where x1 is the prediction score of normal CT and x2 is the prediction score of abnormal CT; the label of the CT image is y=Y, Y∈(0,1), and the label is converted into a hot encoding form, that is, y=[y1,y2]; the classification results x and y of the model inference are input into the cross entropy loss function to calculate the training loss, and the calculation formula is loss=-(y1logx1+y2logx2).

[0037] Step S207: The training loss calculated in step S206 is back-propagated to obtain the optimized gradient of each weight of the model, thereby optimizing the model parameters.

[0038] Step S208: Determine whether the model has converged. If it has converged to a near global optimum, save the model and proceed to step S209. If it has not converged, return to step S206 and iterate and optimize the model.

[0039] Step S209: The trained and saved neural network classification model is inferred on the divided temporal bone high-resolution CT-based imaging lesion classification test set, the classification results of the lesions are output, and visual verification is performed.

[0040] Step S2010: A team of professional clinical physicians evaluates the output lesion classification results to verify the accuracy and generalization ability of the model, and further applies the neural network-based temporal bone high-resolution CT imaging lesion classification model to the clinical diagnosis of lesions.

[0041] This paper extracts specific imaging features from high-resolution temporal bone CT scans to construct a CT feature library for ear diseases, and then develops an ear disease classification system based on these features. This system builds on existing ear disease features from temporal bone CT scans, is based on the inherent logic of recognized imaging knowledge, and is integrated with a deep learning-based automated diagnostic model, resulting in improved diagnostic efficiency in clinical practice.

[0042] References

[0043] 【1】Eroglu O, Eroglu Y, Yildirim M, Karlidag T, Cinar A, Akyigit A(2022) Is it useful to use computerized tomography image-based artificialintelligence modeling in the differential diagnosis of chronic otitis mediawith and without cholesteatoma? Am J Otolaryngol , 43(3): 103395.

[0044] 【2】Wang YM, Li Y, Cheng YS, He ZY, Yang JM, Xu JH (2020) DeepLearning in Automated Region Proposal and Diagnosis of Chronic Otitis MediaBased on Computed Tomography. Ear Hear , 41(3): 669-677.

[0045] 【3】Li X, Gong Z, Yin H, Zhang H, Wang Z, Zhuo L (2020) A 3D deepsupervised densely network for small organs of human temporal bonesegmentation in CT images. Neural Network , 124(75-85).

[0046] The above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. Ordinary technicians in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention. The scope of patent protection of the present invention should be defined by the claims.

Claims

1. Develop an ear disease classification system based on temporal bone high-resolution CT lesion features, characterized by: The following steps are involved: Step S201: The clinical physician team is trained on the use of annotation software, the imaging lesion features of temporal bone high-resolution CT, and the feature annotation standards. The annotation content of temporal bone high-resolution CT includes the disease name and the lesion features of each disease in the temporal bone high-resolution CT, thereby constructing a temporal bone high-resolution CT imaging lesion feature library; Step S202: constructing a classification dataset of ear diseases based on the imaging lesion feature library and disease name labels of the temporal bone high-resolution CT imaging lesions; Step S203: Convert the classification dataset format of the constructed temporal bone high-resolution CT-based imaging lesion classification dataset according to the lesion category, and divide the images and label files in the training set and test set into a ratio of 7:3, that is, 70% of the images are used for neural network model training, and 30% of the images are used for neural network model effect verification and evaluation; Step S204: Load the pre-trained weights of the neural network-based classification model and set hyperparameters, including learning rate, batch data size, input image size, weight decay, and random seed parameters; Step S205: Load the divided temporal bone high-resolution CT-based imaging lesion classification dataset and perform a certain degree of data augmentation on the dataset, including random cropping, upside-down flipping, and left-right flipping strategies, to accelerate model convergence and improve the robustness and generalization ability of the model; Step S206: Input the loaded data into the loaded neural network-based temporal bone high-resolution CT imaging lesion classification model; temporal bone high-resolution CT imaging lesion classification is a binary classification problem, and the model inference obtains a classification result x=[x1,x2], where x1 is the predicted score of a normal CT image and x2 is the predicted score of an abnormal CT image; the label of the CT image is y=Y, Y∈(0,1), and the label is converted into a hot encoding form, i.e., y=[y1,y2]; the classification results x and y obtained by model inference are input into the cross entropy loss function to calculate the training loss, and the calculation formula is loss=-(y1logx1+y2logx2); Step S207: The training loss calculated in step S206 is back-propagated to obtain the optimized gradient of each weight of the model, thereby optimizing the model parameters. Step S208: Determine whether the model has converged. If it has converged to a near global optimum, save the model and proceed to step S209. If it has not converged, return to step S206 and iterate and optimize the model. Step S209: performing inference on the trained and saved neural network classification model on the divided temporal bone high-resolution CT-based imaging lesion classification test set, outputting the lesion classification results, and performing visual verification; Step S2010: A team of professional clinical physicians evaluates the output lesion classification results to verify the accuracy and generalization ability of the model, and further applies the neural network-based temporal bone high-resolution CT imaging lesion classification model to the clinical diagnosis of lesions.

2. The ear disease classification system based on lesion features in temporal bone high-resolution CT according to claim 1, characterized in that: The specific steps of the random cropping are as follows: first scale the input image to 512x512, then crop a 384x384 image block at any position as the input of the model; in the data enhancement of upside-down and left-right flipping, the flipping probability is set to 0.

5.

3. The ear disease classification system based on temporal bone high-resolution CT lesion features according to claim 1, characterized in that: The pathological features include air-fluid level in secretory otitis media, enlargement of the tympanic sinus entrance in cholesteatoma, enlargement of Prussak's space, displacement and erosion of the ossicular chain, enlargement of the mastoid sinus filled with soft tissue and bone destruction, destruction of the scutum, destruction of the facial nerve canal and destruction of the lateral semicircular canal.

4. The ear disease classification system based on temporal bone high-resolution CT lesion features according to claim 1, characterized in that: The annotated disease names include external auditory canal cholesteatoma, middle ear cholesteatoma, suppurative otitis media, cholesterol granuloma, secretory otitis media, temporal bone cancer, cochlear malformation, glomus jugularis tumor and acoustic neuroma.

Images